# Model Configuration
model_name: /large-storage/model/deberta-v3-large # Hugging Face model
labels_encoder: null
name: "span level gliner"
max_width: 12
hidden_size: 512
dropout: 0.4
fine_tune: true
subtoken_pooling: first
span_mode: markerV0

# Training Parameters
num_steps: 10000
train_batch_size: 8
eval_every: 1000
warmup_ratio: 0.1
scheduler_type: "cosine"

# loss function
loss_alpha: -1 # focal loss alpha, if -1, no focal loss
loss_gamma: 0 # focal loss gamma, if 0, no focal loss
label_smoothing: 0
loss_reduction: "sum"

# Learning Rate and weight decay Configuration
lr_encoder: 1e-5
lr_others: 5e-5
weight_decay_encoder: 0.01
weight_decay_other: 0.01

max_grad_norm: 1.0

# Directory Paths
root_dir: span_gliner_logs
train_data: "data/pilener_train.json" # see https://github.com/urchade/GLiNER/tree/main/data
val_data_dir: "/workspace/GNER/regen_data/data/IE_INSTRUCTIONS/NER copy"
# "NER_datasets": val data from the paper can be obtained from "https://drive.google.com/file/d/1T-5IbocGka35I7X3CE6yKe5N_Xg2lVKT/view"

# Pretrained Model Path
# Use "none" if no pretrained model is being used
prev_path: null

save_total_limit: 10 #maximum amount of checkpoints to save

# Advanced Training Settings
size_sup: -1
max_types: 25
shuffle_types: true
random_drop: true
max_neg_type_ratio: 1
max_len: 512
freeze_token_rep: false